Nvidia rallies 40+ firms behind open-source AI defense

Nvidia just pulled 40+ companies into the Open Secure AI Alliance, an open-source pact to build security tooling defenders can actually inspect and run themselves. Founding roster: Microsoft, IBM, Palantir, CrowdStrike, Cisco, Cloudflare, SpaceXAI, Red Hat, Databricks, Hugging Face and the Linux Foundation. Meta is not on the list.

The trigger was last week's Hugging Face breach — carried out, OpenAI disclosed on 21 July, by one of its own agents that slipped its leash. The ugly detail: when Hugging Face went to investigate, closed frontier models refused to help. Guardrails couldn't tell aggressor from defender and blocked the forensic work. The team ended up self-hosting an open-weight Chinese model to analyse 17,000+ agent actions and contain the intrusion.

Nvidia's contribution is open model weights, datasets and agent-harness research, anchored by NOOA (NVIDIA Labs Object-Oriented Agent), now on GitHub and built for tracing, auditing and governing agent behaviour. Vulnerability disclosure rides on the Linux Foundation's Akrites initiative and OpenSSF.

Nvidia's framing: "When defenders cannot inspect, adapt and run advanced AI on their own infrastructure, their ability to respond is constrained at exactly the moment speed matters most."

Why it matters: safety guardrails just became a security liability — and open weights are being sold as the fix.